[snowkide]

Client-side WebSocket traffic is only partly visible in metrics: calls that reach the network are counted in erpc_network_request_received_total indistinguishably from HTTP, and calls rejected before the network (invalid JSON, disallowed method, auth failure, consumer rate limit, network unavailable) are not counted at all. The subscription pressure metrics from erpc#1255 cover connections, subscriptions and notifications per network, but not per client, and not why connections close.

This builds on erpc#1255 instead of adding parallel metrics:

• transport label (http|ws) on erpc_network_request_received_total, recorded on NormalizedRequest (defaults to http).
• erpc_websocket_connections_active gains a network label. The connection's network label is now set once at upgrade instead of on the first subscription, so connections without subscriptions are attributed too.
• erpc_websocket_connections_closed_total{project,network,close_code,initiator}. close_code is bounded to the RFC 6455 range (1000-1015, else other).
• erpc_websocket_requests_total{project,network,category,user,agent_name,outcome}: every JSON-RPC request on a client connection (batch items individually), including pre-network rejections. Outcomes: ok, error, rate_limited, invalid, invalid_batch, method_not_allowed, unauthorized, network_unavailable. category stays n/a until the request is authenticated, so unauthenticated clients cannot create series with arbitrary method names.
• user and agent_name on erpc_websocket_notifications_sent_total and erpc_websocket_subscription_notifications_dropped_total, which become labeled counters. The labels are captured at eth_subscribe time and carried on the subscription (wsclient.SubscriptionLabels).

Notes for review:

• Every new per-client label is on a labeled counter, so deployments that do not want per-user series drop user/agent_name with the existing metrics.customizations label rules. Gauges carry no user label.
• The network label is read from the project config (the value Network.Label() returns). Resolving it with GetNetwork would block the upgrade until a cold network is bootstrapped.
• getOrCreateConn no longer sets networkLabel; it is written once before the connection is shared, so reads need no lock.
• No new latency histogram: per-user latency is already available from erpc_network_request_duration_seconds{user}.
• Docs: Metrics section in operation/websocket, rows in reference/metrics, and the transport label wherever network_request_received_total labels are listed.

Verification

Based on erpc/erpc main @ 094b912.

• go build ./..., go vet ./... and gofmt -l on touched files are clean.
• make test-fast (what CI runs): passes, exit 0, no failing test.
• New/updated tests:
  • TestWebSocket_RequestAccounting: ok, method_not_allowed, invalid and invalid_batch outcomes; transport="ws" on received; connection gauge up and back down; client-initiated close counted with code 1000.
  • TestWebSocket_RequestAccountingAuthRateLimited: the rate_limited outcome.
  • TestWebSocket_CloseCodeLabelIsBounded.
  • TestWebSocket_LogsSubscribeOnAllUpstreamsAndDedupe: the sent counter now asserts the user/agent_name labels.
  • TestNormalizedRequest_TransportDefaultsToHttp.

AI assistance

• Model: anthropic/claude-opus-5.5
• Harness / skills: Claude Code
• Human-reviewed: no, pending

🤖 Generated with Claude Code